Answers not related to facts about kernel-memory

[vvdb-architecture]

I noticed the same thing, using the file-based SimpleVectorDb .

The documents used as "ground facts" are the ones being "closest" to your query, where "closest" is defined using cosine similarity. I found that this kind of similarity often retrieves documents that have nothing to do with the subject at hand.

For example, queries like "How do I do X?" will retrieve documents where X (or something related to X) doesn't even occur.

I am investigating why this is. But this of course never happens with keyword search. Documents containing X would be retrieved by definition.

This is all very strange. The investigation continues.

You can find out which facts are being selected by placing a breakpoint at the call of SearchClient.GenerateAnswerAsyncin SearchClient.

[rosieks]

Actually in my case facts are correct but OpenAI ignores them completely.